Duplicate Video Detection Based On Deep Learning

With the development of deep learning, many problems which cannot be solved by traditional methods are constantly being overcome. In multimedia information processing, like image voice and text, it becomes a trend to combine deep learning technology with high performance computing. In the Information Age, video as an important carrier, plays a huge role in disseminating information. The information expansion brings new challenges to multimedia video retrieval. Duplicate video detection can reduce the redundant information in video database, also plays an important role in copyright protection.Content based duplicate video retrieval could be divided into three parts:1. Key frame extraction.2. Feature extraction and indexing.3. Matching and filtering. Current methods mostly use local features of image. With the development of deep learning, convolution neural network has achieved good results in image recognition and retrieval, but has not been widely applied to video retrieval and related areas.The main work of this paper are divided into two parts, one is about frame extraction, the other is about duplicate detection. In key frame extraction, we proposed a fast method based on shot boundary detection and a robust method based on clustering, it works well for short video. In duplicate video detection, we designed and completed the experiments in three parts:key frame extraction, image feature selection and matching sorting strategy. All the videos used in experiments based on CC WEB VIDEO database. From the experiments results, the key frame selection, can significantly reduce the retrieval time and space overheads; the detection accuracy for CNN feature is significantly higher than SIFT feature, when the recall is 0.8, the precision is no lower than 0.97; Matching strategy using key frame weights can improve the rank performance in a certain degree.